| 35 | namespace { |
| 36 | |
| 37 | std::string GetAddTableCode(int src_count) { |
| 38 | std::string code = R"( |
| 39 | #include <metal_stdlib> |
| 40 | using namespace metal; |
| 41 | |
| 42 | struct uniforms { |
| 43 | int4 src_size; |
| 44 | }; |
| 45 | |
| 46 | $0 |
| 47 | kernel void ComputeFunction( |
| 48 | $1 |
| 49 | uint3 gid[[thread_position_in_grid]]) { |
| 50 | if (static_cast<int>(gid.x) >= params.src_size.x || |
| 51 | static_cast<int>(gid.y) >= params.src_size.y) { |
| 52 | return; |
| 53 | } |
| 54 | |
| 55 | FLT4 value = FLT4(0.0f); |
| 56 | int linear_index = (int(gid.z) * params.src_size.y + int(gid.y)) * |
| 57 | params.src_size.x + int(gid.x); |
| 58 | )"; |
| 59 | for (int i = 0; i < src_count; ++i) { |
| 60 | code += " value += src_buffer" + std::to_string(i) + "[linear_index];\n"; |
| 61 | } |
| 62 | code += " $2\n"; |
| 63 | code += " dst_buffer[linear_index] = value;\n"; |
| 64 | code += "}\n"; |
| 65 | return code; |
| 66 | } |
| 67 | } // namespace |
| 68 | |
| 69 | std::vector<ComputeTaskDescriptorPtr> Add(int id, |